The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses git commands (diff, status, rev-parse, ls-files) to inspect the repository's state and file changes. These are standard operations for a security auditing tool and do not involve executing untrusted binaries or scripts.
[INDIRECT_PROMPT_INJECTION]: This skill interacts with untrusted data in the form of repository filenames and code diffs. It proactively mitigates injection risks by using NUL-terminated output (git -z) and a Python script to JSON-escape filenames before display, ensuring that malicious filenames cannot execute terminal control sequences or mislead the agent.
Ingestion points: Git command output and file content from the local repository.
Boundary markers: Uses NUL-separation and retrieval boundaries to isolate different sources of changes.
Capability inventory: Subprocess execution of git and file read operations.
Sanitization: Implements a Python-based NUL-aware parser that serializes data into JSON format to ensure safe rendering in the agent's context.
[DYNAMIC_EXECUTION]: A Python snippet is used to parse stdin and format output as JSON. The snippet is statically defined within the skill instructions, uses only the standard library (json, os, sys), and serves a purely structural purpose for data processing.

security-review